The bar we set for Ireland, and why it just cleared

On Saturday we published a piece called "Galway starts Monday and our model will sit it out." By the small hours of Sunday morning, less than a day later, we had taken it down.

Nothing sneaky happened in between. The piece was honest when we wrote it. Irish predictions were switched off, waiting on a backtest that could only run on the machine that trains the model, and that machine had been unreachable. Then it became reachable. The backtest ran that same evening, on 3,117 Irish races already sitting in our database, and it cleared every one of the four bars we had written down in advance.

It is no longer a promise. Galway opened this morning and today's card went out with Irish races priced alongside the British ones for the first time.

We are telling you this rather than quietly swapping one post for another, because the swap is the more interesting story. Whether a model is calibrated on a type of race it has not been tested on is either measured or it is not. The honest version of "we do not know yet" is to publish nothing until the number exists, then publish exactly what the number said, whichever way it landed. It could just as easily have failed. It did not.

What follows is the actual test and the actual result, so you can judge it as evidence rather than take our word for the announcement.

The bar it had to clear

The gate was four numbers, fixed before the test ran, precisely so nobody could move them afterwards to fit the answer.

The test scored British and Irish racing side by side over the same window, so every comparison below is like for like. That matters more than it sounds. Our published track record runs over a longer period, so setting an Irish backtest figure against the live British one would compare two different populations, and it would flatter Ireland.

• Sample. At least 300 resolved Irish races, ideally more than 500. The backtest scored 3,117, alongside 10,701 British races in the same window.
• Calibration. The gap between the model's Brier score and the market's, on Irish races only, no worse than about +0.020. A Brier score measures how well a set of probabilities matched what actually happened, lower is sharper. The Irish gap came out at +0.0042, well inside the limit, and narrower than the same model's gap on British racing over the same window, +0.0063. On the one number this gate exists to check, Irish racing measured slightly more honest than the British racing we already publish, not less.
• Return. Irish return to starting price had to sit inside the band British racing already occupies. It came out at −11.1% to SP, against −11.3% for Britain over the same races. Both are losses of much the same size.
• Strike. Irish top-pick strike rate somewhere in the 20 to 26 per cent range. It landed at 25.4%, against a British 27.2%.

Four for four. That is the entire basis for the switch. Not a hunch that Irish racing would behave the same way as British racing, a measurement that it did.

For context rather than comparison, the live British record we publish and update every night currently sits at −12.2% across 7,855 races. It covers a longer period than the backtest above, which is why the two British figures differ, and why we have not set them against each other.

What clearing the bar does not mean

Better calibrated is not the same as profitable. −11.1% to SP on flat stakes is still a heavy loss, roughly the bookmaker's margin on the race, and nothing in this backtest says an Irish card is a good bet because our model rates a horse highly. The gate was only ever a claim about honesty: that an Irish probability from this model deserves the same trust as a British one. It says nothing about whether either one makes money, because neither reliably does. Of the 24 staking and selection systems we have tested across 28,098 real races, 0 sit in profit.

Galway's cards now carry the same model overlay every British card gets, and get scored the same way the next morning, wins and losses both published. For the racing itself, our Galway Festival guide and results hub cover the week as it happens. To watch the model's own record rather than take this piece's word for it, start at the track record or the Silicon Tipster League, which now has an Irish opinion to show alongside the British one.